Robust Hierarchical Grouping Learning Immune to Missing Data for Voltage Stability Assessment

This paper proposes an innovative long-term voltage stability assessment (VSA) approach tolerant to missing data based on hierarchical grouping convolutional neural network (HGCNN). The proposed HGCNN is a kind of enhanced convolutional neural network (CNN) incorporating multi-round feature grouping and integration part, able to achieve high accuracy of VSA with and without missing data, where missing data may be the loss of lots of data points from multiple measurement sensors. In the proposed approach, each measurement series is independently transformed into a Gramian angular field (GAF) matrix as features at first. Subsequently, a feature grouping strategy clusters these GAF features according to the system topology, and then numerous CNN blocks are employed to filter these features and generate abstract representations. Furthermore, multiple abstract representations within the same group are then integrated using a symmetrical function. Finally, after multiple rounds of feature grouping, abstract representation extraction by CNN block and feature aggregation, an output block of CNN is employed to perform voltage stability margin (VSM) estimation and VSA status classification. Numerous experimental evaluations and comparisons show that the proposed approach remains highly accurate and robust with and without the presence of missing data.